Klitzke, L., Koch, C.
Abstract:
In this paper, a novel approach is presented for intrusion detection in the field of wide-area outdoor surveillance such as construction site monitoring, using a rotatable stereo camera system combined with a multi-pose object segmentation process.
In many current surveillance applications, monocular cameras are used which are sensitive to illumination changes or shadow casts. Additionally, the object classification, spatial measurement and localization using the 2D projection of the 3D world is ambiguous. Hence, a stereo camera is used to calculate a 3D point cloud of the scenery which is nearly unaffected by illumination changes, therefore enabling robust object detection and localization in the 3D space. The limited viewing range of the stereo camera is expanded by mounting it onto a rotatable tripod. To detect objects in different poses of the camera, pose-specific Gaussian Mixture Models are used. However, changing illumination outside the current field of view of the camera or spontaneously changing lighting conditions caused by e.g. lights controlled by motion sensors, would lead to false-positives in the segmentation process if using the brightness values. Hence, segmentation is performed using the calculated point cloud which is demonstrated to be robust against changing illumination and shadow casts by comparing the results of the proposed method with other state of the art segmentation methods using a database of self captured images of a public outdoor area.